JP2580826B2 - Feedback neuron model - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a feedback neuron model, and more particularly to a feedback neuron capable of processing time-varying information with a neural network modeling a brain function. Related to cell model.

[Conventional technology]

Neuronal models that model brain functions have recently become well known.

As shown in FIG. 3, a conventional neuron model uses a weighted signal Wi, j, Oi (t) obtained by weighting an input signal Oi (t) (i = 1 to n _K ) from another group of neurons. ) And the weighted signal are added, and in the cell internal state And an axon mound 3 that converts the cell internal state Xj (t) into a cell output value Oj (t).

As described above, the conventional neuron model uses a linear sum of input signals Oi (t) from other neurons with weights (Wi, j). Is the cell internal state Xj (t), and a sigmoid function f (Xj (t)) = 1 / (1 + exp (−Xj) as an output conversion function.
(T)) was used to determine the output value Oj (t).

[Problems to be solved by the invention]

Since the output value Oj (t) of this conventional nerve cell model is determined only by the input signal Oi (t) from another nerve cell at the current time t, the temporal change of the input signal Oi (t) is Not used at all within. In other words,
No matter what input signal Oi (t) is present in the past, it has no effect on the output of the cell. On the other hand, the human brain
As can be seen from examples of sight and hearing, advanced information processing is performed by capturing not only instantaneous signals but also their temporal changes. Therefore, in the conventional neuron model, the output value Oj (t) is the input signal Oi as the brain neuron model.
There is a disadvantage that it is not affected by the temporal change of (t).

[Means for solving the problem]

The feedback neuron model of the present invention is a synapse connection unit that receives a signal obtained by weighting the output signal of another group of nerve cells, and the weighted signal input from the synapse connection unit is added to the internal state of the cell. In a nerve cell model having a cell body part that generates a cell and an axon hill that converts the internal state of the cell into a cell output value, the cell body part of the cell itself has the internal state of the cell at the current time Xj.
(T) is also affected by the internal state of the past cell of the cell body part of the self.

[Action]

The feedback neuron model according to the present invention uses the cell internal state Xj (t) one hour before the cell internal state Xj (t) to be generated in the cell body so as to depend on the cell output value, in other words, the past input signal to the cell. A feedback coupling for feeding back the state Xj (t-1) is provided. In this case, as the feedback junction coefficient Xj is closer to 1, the cell output value Oj (t) also depends on a past input signal in which the cell output value Oj (t) is farther, but the cell internal state Xj (t) does not diverge. Therefore, the magnitude of the feedback joining coefficient aj is set to less than 1.

〔Example〕

 Next, the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of a feedback neuron model according to the present invention, in which a synapse connection section 1 receiving an input signal from another nerve cell group and a cell body section generating a cell internal state at time t 2, an axon mound 3 for exchanging the cell internal state with a cell output value, and a feedback connection 4 for feeding back the cell internal state one time ago. The feedback coupling described above establishes a neuron model that also depends on past input signals to the cell.

FIG. 2 is a block diagram of an embodiment of a hierarchical neural network constructed using the feedback neural cell model of FIG.

In FIG. 2, the input signal i (t) is a scalar, but there is no particular problem in using a multidimensional vector. The neural network is composed of M layers ranging from an input layer 11 as a first layer to an output layer 15 as an Mth layer.
(K = 1 to M) cells are included. The connection between nerve cells is performed only between adjacent layers, and there is no connection between layers. The signal input signal i (t) input to the input layer 11
Is the information on the past input signal in each nerve cell, and in this embodiment, the cell internal state Xj (t-1) one time ago.
Nakara is adding a unidirectionally flowing the neural circuit network towards the input layer 11 to the output layer _{^{15, ▲ O M j ▼ (}} t) (j = 1~
nM) is output from the output layer 15.

In FIG. 2, 101 included in each layer indicates cells,
Reference numeral 102 denotes a feedback connection. Note that a straight arrow indicates a bond.

Intracellular to indicate the symbol ▲ X ^K _j ▼ (t) is inside the cell state of the K layer j cells at time t, e.g. ▲ X ² ₁ ▼ (t)
Is the state inside the cell at time t of the second layer first cell. The symbol ▲ O ^K _j ▼ (t) is a K layer J at time t
The output value of a cell, e.g. ▲ O ² ₁ ▼ (t) is first second layer
This is the cell output value at time t of the cell. The symbol ▲
a ^K _j ▼ is set to 1 or less so that the cells internal state does not diverge in the feedback coupling coefficient of the K layer J cells, for example ▲ a ² ₁ ▼ is a feedback coupling coefficient in the first cell the second layer. The symbol ▲ W ^{K, K + 1} _{i, j} ▼
The weight of the connection between the i-th cell in the Kth layer and the jth cell in the (K + 1) th layer is shown. The symbol θθ ^{K + 1} _j ▼ indicates a threshold value required for a stable network configuration of the cell internal state in the K + 1 layer J cell, and the cell internal state in the second to K + 1 layers is −1. Based on network construction requirements.

In this manner, the input signal i (t) is output by the present neural network in consideration of the information on the past input signal and the output value O (t).
(J = 1 to nM). Coupling factor of neural network ▲
The learning of W ^{K, K + 1} _{i, j} ▼ and the feedback coupling coefficient Ka ^{K + 1} _j ▼ can be effectively performed by slightly modifying the well-known error back propagation method.

In this way, by feeding back the state for one time to each cell, time-series information in the network is confined, and processing corresponding to time-varying of the input signal becomes possible.

〔The invention's effect〕

As described above, according to the present invention, it is possible to process a temporal change of an input signal by configuring a hierarchical neural network using a neural cell model to which a feedback mechanism of a cell internal state one time ago is added. This has the effect that advanced information processing closer to the human brain can be performed efficiently.

[Brief description of the drawings]

FIG. 1 is a block diagram of one embodiment of a feedback neuron model of the present invention, FIG. 2 is a block diagram of one embodiment of a hierarchical neural network configured using the feedback neuron model of FIG. FIG. 3 is a block diagram of a conventional nerve cell model. 1 ... synaptic connection part, 2 ... cell body part, 3 ... axon mound, 4 ... feedback connection, 11 ... input layer, 12 ...
2nd layer, 13... Kth layer, 14... K + 1th layer, 15... Output layer, 101... Cell, 102.

Claims (2)

(57) [Claims] 1. A cell for generating a cell internal state by adding a weighted signal input from the synapse connection section and a synapse connection section for receiving a signal obtained by weighting an output signal of another nerve cell group. In a neuron model having a body part and an axon hill that converts the internal state of the cell into a cell output value, when the cell body part of the cell itself assumes the internal state of the cell at the current time as Xj (t), A feedback neuron model characterized by being affected by the past internal state of cells in its own cell body. 2. The cell internal state Xj to be generated by the cell body part
(T) is a signal obtained by weighting the output signals of the other nerve cell groups and the cell internal state Xj (t−1) one time earlier by 1
The feedback neuron model according to claim 1, wherein the feedback neuron model is generated by adding the following weighted signal.